Machine learning-based thermokarst landslide susceptibility modeling across the permafrost region on the Qinghai-Tibet Plateau

Abstract

Thermokarst landslides (TL) caused by the thaw of ground ice in permafrost slopes are increasing on the Qinghai-Tibet Plateau (QTP), but the understanding of the spatially suitable environmental conditions including terrains and climate for them has not been fully established. Here, we applied multiple machine learning models and their ensemble to explore factors controlling the TL and map its susceptibility at a fine resolution. The models were calibrated and validated using a split-sample approach based on an inventory of TLs from the remote sensing data. The models indicated that summer air temperature and rainfall were the most two important factors controlling the occurrence and distribution of TLs, provided that other geomorphic conditions (i.e., slope, solar radiation, and fine soil) were suitable. The final ensemble susceptibility map based on downscaled climate data and terrain data suggested that ca. 1.4% of the QTP land was classified in high- to very high-susceptibility zone, which is likely to increase in response to future climate change. This study integrated local topography and climate in susceptibility modeling and provided new insights into the geomorphic sensitivity to climate change but also the engineering support over the QTP.